Skip to content

The engine track

This track is how an engine whose planes are measured in code metrics became the fastest route to every measured accuracy at 16M rows on one GPU. It teaches that through the episodes that did it.

What this is

The algorithm track teaches how gradient boosting works. This track teaches how bonsai got fast. It is HPC pedagogy told case-method, like a college class. Each chapter is one real engineering episode from the project record, built up slowly, with the discipline as the through-line. The losses and the refutations are the curriculum, not the footnotes.

Who it is for

A developer who knows some C++ and has never priced a GPU kernel. You do not need CUDA experience. You need to be willing to trust a measurement over an intuition, which is the entire method.

The method in brief

Four rules recur across every case.

  • Instrument first. No optimization begins until instrumentation has decomposed and priced the cost it attacks.
  • Price before building. An edge move states its model price from measured constants before anyone writes a kernel.
  • Same-pod discipline. Two identical GPUs measured 25% apart across the fleet, so every delta is a same-pod delta.
  • Refutation is a deliverable. A measured no, with the conditions that would reopen it, is worth as much as a win.

The chapters

  • E1. The marginal round. Instrumentation cancels a kernel rewrite, then a price list cuts the 16M round from 155 to 104 ms.
  • E2. The missing bin. An acceptance test fails, and the fix closes a train/predict skew hiding in every fitted model.
  • E3. The parity verdict. A data-parallel multi-GPU engine, built and measured to parity, then parked as an experiment.
  • E4. The resident objective. Deleting the per-tree host round-trip, and the round falls again from 104 to 64 ms.
  • E5. The ceiling. 500M rows by 100 features trained end to end on one 80GB card.
  • E6. The wide-data wall. A production field report at 16k features, the cache arithmetic behind a 2-6x cliff, two wrong theories killed by an interleaved A/B, and the tiled layout that dissolves the trade-off.